Authored by Dr. Mikel J. Harry, Six Sigma Management Institute, Inc.
Quite often, Lean Six Sigma Practitioners and University Professors send me a real-world project related to something they are currently working on. As would be expected, they frequently ask me to share my thoughts about what they are doing, how they are doing it or the conclusions they are deriving from the data. Sometimes the questions are about process control, or of a diagnostic concern. At other times, their questions relate to a statistically designed experiment.
In this context, its likely that novices and practitioners of Lean Six Sigma would want to know about such problems and how those problems can be approached, structured, analyzed and successfully resolved. To do this, we would need a common foundation for discussion. If such a foundation could be created, highly insightful discussions would likely emerge. In other words, by using an actual case study to fuel the dialogue, such discussion-centric objectives can be realized.
In response to this idea, I decided to post a recent communication that came to my attention by way of an email (see below). The correspondence was from a university professor in India. The situation involved a live experiment (full factorial design) that was executed as a class project. A quick overview of the issue, driving questions and raw data were provided in the email (see below). The professor’s identity and email have been withheld.
The full data set is available as a spreadsheet download at: Raw Data Set for Experimental Project
After downloading the raw data, I began my usual procedure of “playing around with the numbers” so as to “get a feel for the situation,” as well as contemplate those things that might prove to be an issue during the course of a formal analysis. Of course, this procedure is more exploratory in nature than diagnostic.
For example, I often start by graphing the data in different ways — looking for any “oddities” that may be present. If any such peculiarities surface, I might choose to dig a little deeper using classical statistics. As yet another example, if there is an issue with the residuals (lack of fit to a normal probability distribution), I will often replace the suspect values with the mean (assuming there are only a few such oddities) and then repeat the analysis. In this way, I can get a feel for how robust certain analytical tools will be to a violation of the assumption that the residuals fit a normal distribution. If the impact of such replacement is negligible, I run with the original data. If not, more digging into the matter would likely be warrented.
With such preliminary analyses out of the way (and assuming a green light), I move into a more formal analysis of the data, but done so in a highly deductive manner. Of course, we could write a book on any of the latter points; consequently, the reader should recognize that such “ways and means” are quite general in nature and in no way should be treated as a formal process of investigation. They are simply a part of my personal trade-craft developed over the last 30+ years of practice.
Meanwhile back at the ranch, my response to the kind professor is provided below.
My response letter is available as a PDF download at: Dr. Harry’s Response to University Professor.
In summary, it should be fairly evident that even the most rudimentary analyses of a statistical nature under the best of conditions can open the door to the potential for “analysis paralysis.” This means that a statistical analysis of empirical data, especially that resulting from a designed experiment, can become quite protracted when attempting to test and validate all of the underlying assumptions and prerequisite conditions.
Analogously speaking, the phenomenon of analysis paralysis is much like setting a goal to explore the Amazon river while attempting to investigate all of the little tributaries along the way. Essentially, there are so many tributaries, your goal of exploring the river gets lost in all the sub-explorations that seem interesting or necessary at the time, so to speak.
In the final analysis, this article addresses the need to “keep it simple and on-course,” but paradoxically, doing so comes with risk. However, this practitioner believes that its far better to complete 100 experiments with 80% confidence than 10 experiments with 100% assurance. I believe an organization’s shareholders would agree. The key is to maintain forward momentum and not get hung-up in all the theoretical details.
In my experience, those organizations that subscribe to this way of thinking laugh all the way to bank. While this philosophy will likely make the “purists” shake in their boots, they often forget that seven digits of precision in a two digit world don’t get the bills paid and the dividends distributed among the shareholders.
Business Phone: 480.515.0890
Business Email: Mikel.Harry@SS-MI.com
Copyright 2013 Dr. Mikel J. Harry, Ltd.